Sam Cheyette

Dream team: Tracey Mills (@traceym.bsky.social), Nicole Coates, Alessandra Silva (@alessandra-silva.bsky.social), Kaylee Ji, Steve Ferrigno (@sferrigno.bsky.social), Laura Schulz, Josh Tenenbaum.

Our results highlight program learning as a powerful, potentially distinctive, and early-emerging ability that humans deploy to learn structure. Our comparative/developmental results also raise many exciting questions—check out our paper for those + some speculations :). osf.io/preprints/ps...

Distribution of log likehoods (y-axis) of data by sequence in each group (x-axis), under each of the learning models. Log likelihoods are averaged across participants (and trials, for monkeys) within each sequence. Monkeys and 3-yo are both mostly best fit by linear/ linear+previous point models, older children and adults are overall best fit by the LoT model.

We compared various Bayesian learning models with each population. The main takeaway: children as young as 4-years-old showed adult-like program induction on our task. Monkeys and 3-year-olds mostly used local extrapolation.

Scatter plots of accuracy across different pattern types for humans (x-axis) and monkeys (y-axis), split by age (youngest to oldest in each panel).

In contrast, despite extensive training on algorithmic patterns, the monkeys relied on a simpler local linear extrapolation to make predictions. Interestingly, 3-year-olds mostly used this same strategy—and their accuracy across patterns correlated with monkeys much more than with adults.

Consistent with a program-learning account, older children and adults' initial predictions typically show early multimodal uncertainty, but converge on the true pattern after only a handful of observations.

Predictions for how representative spatiotemporal patterns will continue, at a selected set of timepoints, generated by participants in four different population: monkeys, 3 year-old children, 4-7 year-old children, and adults. For adults and children, each dot represents the prediction of one participant. Older children and adults show more structured and often multimodal predictions, whereas 3-year-olds’ and monkeys’ predictions tend to track the locally linear trend.

We found striking differences across both development and species in our task. Below are some examples of predicted continuations on various sequences in each population.

Illustration of program learning as implemented by the LoT model. Starting at the bottom left, the learner observes the partially revealed pattern, then computes a distribution over generative programs conditioned on this observation, and finally runs the programs forward to extrapolate the pattern and predict the next point. Predictions are weighted by the posterior probability of their generative program. Programs are drawn from a grammar containing compositional functions and domain-specific motor and geometry primitives.

If participants are learning structured programs in an expressive “Language of Thought”, they should (1) be able to learn the sequences we tested by the final timepoint; but (2) show patterns of multimodal uncertainty reflective of possible algorithms that are consistent with the sequence so far.

Fig. (A) Task as seen by children and adults. The large star is at the most recently revealed sequence location, with earlier locations indicated by smaller points. Monkeys saw an analogous display of red circles against a white background, with later circles brighter than earlier circles, and the most recently revealed circle larger than the rest. (B) An example of a sequence unfolding over from the third step (left) to the sixth step (right) and predictions made by adults.

One possibility is that we can “learn by programming” rapidly inferring structured algorithms to model our observations.

Our paper tests this ability in adults, 3-7yo children, and two rhesus macaques. Participants predicted how 2D sequences would unfold starting from the first few timepoints.

People seem wired to uncover hidden structure: we pick up the rules of games after a few turns, see figures in clouds and constellations, and riff on songs. What are the computational mechanisms that make this rapid structure learning possible?

Very excited to share a new preprint that’s been brewing for a long time! This work was led by the exceptional @traceym.bsky.social, and made possible by a developmental + comparative + computational dream team.

osf.io/preprints/ps...

Do Humans Use Push‐Down Stacks When Learning or Producing Center‐Embedded Sequences? Complex sequences are ubiquitous in human mental life, structuring representations within many different cognitive domains—natural language, music, mathematics, and logic, to name a few. However, the....

New work with @samcheyette.bsky.social & Susan Carey testing the memory architecture used when learning/producing center-embedded sequences. Adults don't use Push-Down Stacks as is often assumed, instead they rely on a Queue-like memory architecture onlinelibrary.wiley.com/doi/10.1111/...

Dissemination of Erroneous Research Findings and Subsequent Retraction in High-Circulation Newspapers: A Case Study of Alleged MDMA-Induced Dopaminergic Neurotoxicity in Primates Ensuring the public is informed of retractions has proven difficult for the scientific community. While it is possible that newspapers focus differential attention on publication of scientific arti...

Oh man I came across that years ago, it's nuts - was published in Science too!

www.tandfonline.com/doi/abs/10.1...

a man in a blue shirt and tie is pointing at a woman and saying " too legit to quit " . Alt: a man in a blue shirt and tie is pointing at a woman and saying " too legit to quit " .

Whelp. See you later 1.8 Million in NSF research funds -- all designed to better understand learning mechanisms in early childhood so we can develop effective early childhood educational interventions.

Proud of Harvard for standing up to fascism, though.

We will persist.

How Universities Became So Dependent on the Federal Government For decades, universities got billions in federal dollars for research. The relationship was mutually beneficial, until President Trump decided it wasn’t.

There are so many ways one could provide context for this data.

For example, in the last five years universities have received 52-55% of their research funding from the federal government. That's the lowest percentage since the 1950s. 1/x ncses.nsf.gov/surveys/high...

Six trapezoids.

1. Quick—which of these shapes is different from the others?

The foundations of America’s prosperity are being dismantled Federal scientists warn that Americans could feel the effects of the new administration's devastating cuts for decades to come

For decades, the US government has painstakingly kept American science #1 globally—and every facet of American life has improved because of it. The internet? Flu shot? Ozempic? All grew out of federally-funded research. Now all that's being dismantled. 1/ www.technologyreview.com/2025/02/21/1...

hopefully my first and last ever mildly political post. from here on out it's memes and papers

"this rendition emphasizes her eyes, carefully positioned to create the illusion of a gaze that follows you"

and ascii mona lisa

I am proud to report that humans still have the upper hand on palindromes

Come join the Cognitive Origins Lab at UW-Madison! We are hiring two full time lab managers to start this summer! One specializing in child development and one in non-human primate cognition. Application links below.

Human behavior is hierarchically structured. But what determines *which* hierarchies people use? In a preprint, we run an experiment where people create programs that correspond to hierarchies, finding that people prefer structures with more reuse.

arxiv.org/abs/2311.18644

1/7